The Edge: Seeing Signals Before They Become Prices
Most retail and institutional investors discover emerging technology trends through financial media — by which point the opportunity is already priced in. Innovation Trader reads the primary sources: academic papers, patent filings, government databases, and a continuously updated knowledge base of 54,000+ curated tech articles, identifying companies with genuine breakthrough exposure days or weeks ahead of mainstream coverage.
"The IBM quantum computing $1.4B government contract appeared in our signal database before it was covered by any financial media. The system identified the sector acceleration and increased nuclear/quantum domain weighting automatically."
Core Sleeve (up to 5 positions)
Large-cap companies (>$10B) with proven revenue and direct tech exposure. Capital preservation with 5% stop loss. Rotates when a materially stronger signal arrives. Provides stability while satellite positions run.
Low volatility anchorSatellite Sleeve (up to 7 positions)
Early-stage innovators ($50M–$10B) at the commercialisation inflection point. 10% stop loss allows for volatility. Trailing stops lock in gains as positions appreciate — no arbitrary profit-taking that caps upside.
Asymmetric upsideNine-Gate Signal Discovery Pipeline
Every potential trade passes through a multi-stage validation process that combines quantitative screens, LLM-powered analysis, and geopolitical context before any capital is committed.
Multi-Source Signal Discovery
Daily pre-market scan across 33 emerging technology domains. Sources are weighted by historical performance and reordered each morning based on velocity signals from the proprietary article database.
Supabase Velocity & Sentiment Intelligence
Before the LLM is invoked, the system queries a 54,000+ article database for acceleration signals. Domains where coverage is increasing week-on-week get prioritised. Articles are scored for commercial actionability — "company X deployed Y" outranks "researchers propose Z". This surfaces genuine deployment signals, not just research noise.
Geopolitical & Macro Context Layer
Before scoring any signal, the system loads current macro context from four sources. Political and economic events are automatically mapped to sector multipliers — a ceasefire in the Middle East boosts clean energy scores; rising US/China tech tensions reduce semiconductor allocation.
LLM Company Extraction (GPT-4o / Claude)
Each validated innovation is passed to GPT-4o, which extracts 1–3 publicly traded companies best positioned to benefit, with a confidence score (0–100%). Groq/Llama-3 serves as a high-speed fallback. The prompt is domain-aware and instructed to prioritise small/mid-cap companies with direct exposure over diversified conglomerates.
Quantitative Pre-Filter
Fast quantitative gates applied before the expensive viability check. Eliminates candidates with insufficient liquidity, stocks that have already run significantly, or zero-revenue micro-caps — saving LLM cost and reducing noise.
10-Factor Viability Score — The Investment Committee
This is the core judgement layer — a structured scoring model that asks the same questions a disciplined investment committee would ask before committing capital. Each factor is scored independently; the company must reach 50/100 to proceed. This threshold was deliberately set to be demanding rather than permissive — the goal is to surface the top 1–2% of signals, not to find reasons to trade.
Momentum Confirmation — Timing the Entry
A company can have a genuine breakthrough signal and pass all 10 viability factors — and still be the wrong trade right now. This gate solves the timing problem. It answers a single question: has the market already moved on this signal?
Consolidated Swap Review
All qualified signals are ranked simultaneously against all current positions. The best new signal is matched against the weakest holding. Override signals (acquisitions, FDA approvals, government contracts) bypass the threshold entirely and rotate immediately. Maximum 3 swap proposals per cycle.
Execution & Risk Management
Cash-only enforcement — no margin ever deployed. Sleeve-appropriate position sizing and stop losses. Trailing stops that follow the stock up without capping upside: satellite positions have stops at +10% after a +25% gain, rising to +90% after +150%.
Not Just ML — A Full Agentic Investment System
The system goes beyond a prediction model. It is an autonomous agent with tool use, memory, self-improvement, and natural language interaction.
Agentic Decision Making
The system makes autonomous buy, sell, and swap decisions within defined risk parameters. In confirmation mode, it proposes trades with full investment rationale to the portfolio manager via Telegram, awaiting approval before execution.
Claude Haiku / GPT-4oNatural Language Interface
The portfolio manager interacts with the system in plain English via Telegram. Claude has full access to portfolio data, signal history, article database, and geopolitical context. It can propose and execute trades on instruction.
Anthropic ClaudeClaude-Written Swap Proposals
Each swap proposal is written by Claude as an investment case — sell rationale, buy thesis, risks, historical win rates for the domain and signal type, and Supabase acceleration data. Reads like a brief from an analyst, not a score comparison.
Structured reasoningSelf-Reinforcing Learning Loop
Four-level optimisation running continuously. Exit decisions tracked 30 days post-sale to verify correctness. Source weights, domain multipliers, and entry thresholds auto-adjust based on actual trade outcomes.
Active learningA System That Gets Better Every Week
The Feedback Loop
Most algorithmic systems are static — they run the same logic indefinitely. Innovation Trader continuously measures the accuracy of its own predictions and adjusts its behaviour. It is building a track record that it actively learns from.
| Level | Mechanism | What it adjusts | Activates |
|---|---|---|---|
| L1 Parameter Tuning | Analyses win rates by confidence bucket and exit type | MIN_CONFIDENCE, stop loss thresholds, swap aggressiveness | Daily (10+ trades) |
| L2 Strategy Memo | Claude reviews all performance data weekly | Domain priorities, source weighting guidance, risk posture | Every Sunday |
| L3 Swap Quality | Post-sale tracking — did the sold stock keep falling or recover? | Swap threshold, auto-close aggressiveness | Daily (30-day lag) |
| L4 Feature Weights (RL) | Ridge regression on closed trade features vs actual returns | Confidence, viability, signal type weights in opportunity scoring | 50+ closed trades |
The Data Moat
The system's core advantage is not the algorithm — it is the accumulation of proprietary, structured data that improves signal quality over time.
54,000+ Article Database
Continuously ingesting media and academic articles across all 33 tech domains. Each article is taxonomy-tagged, sentiment-scored, and relevance-ranked. Growing daily — the signal database that exists after 12 months of operation cannot be replicated quickly.
4,600+ Company Assessments
Every company the system has evaluated — with LLM reasoning, viability scores, rejection reasons, and eventual outcomes. This becomes a training dataset for filter calibration and a reference for avoiding repeat mistakes.
Signal Performance History
Every signal logged from discovery to exit, with 30-day post-exit tracking. The system knows which source types, tech domains, and signal patterns produce the best returns in its specific strategy — and adjusts accordingly.
Geopolitical Intelligence Layer
Real-time sector adjustments from Polymarket prediction markets, GDELT global news monitoring, World Bank economic indicators, and curated geopolitical RSS feeds. Political and macro events automatically translated into domain multipliers.
33 Emerging Technology Domains
Focused on technologies at the pre-commercialisation stage — where academic and patent signals appear before equity market pricing.
Institutional-Grade Risk Controls
- Cash-only enforcement — zero margin, never spends more than 95% of available cash
- Sleeve position limits — max 5 core + 7 satellite, enforced at execution
- Sector concentration cap — maximum 4 positions per broad sector
- Earnings calendar filter — never buys within 3 days of earnings (avoids binary event risk)
- Daily circuit breaker — halts all trading if portfolio drops 5% in a single day
- Trailing stops — satellite positions: floor at +10% after +25% gain, rising to +90% after +150%
- Auto-close — positions scoring below threshold closed immediately without waiting for a replacement
- Momentum confirmation — RSI <75 and 5-day gain <15% required before entry
- Override signals — acquisitions and FDA approvals bypass normal thresholds for immediate rotation
- Weekly exit audit — 30-day post-sale tracking verifies whether exit decisions were correct
Production-Grade, Cloud-Native
Cloud Deployment
Running 24/7 on Google Cloud with systemd auto-restart, swap memory, and automatic recovery. Deploys from GitHub — any code change is live in under 60 seconds.
Broker Abstraction
Single interface supports Alpaca, IBKR, Trading212, and Tradier. Currently running paper trading on Alpaca. Switch to live Trading212 is a single environment variable change.
Telegram Control Panel
Full portfolio management from a phone. Commands: /portfolio, /watchlist, /scan, /geo, /macro, /optimizer, /exits, /missed, /snapshot. Natural language chat with full data access and trade execution capability.
Live Dashboard
Auto-updating HTML dashboard published to GitHub Pages daily. Shareable link showing current portfolio, recent trades, discovered signals, and performance statistics.
Where the Alpha Comes From
Innovation Trader is not a momentum strategy, a quant factor model, or a news sentiment trader. It occupies a specific and defensible niche: the gap between scientific publication and equity market pricing.
"Academic papers, patent filings, and government contract databases are public information. But they are written in technical language, published in domain-specific journals, and require expert synthesis to connect to investable companies. The market is slow to do this. We are not."
The Core Inefficiency We Exploit
Markets are broadly efficient at pricing known information. They are much less efficient at pricing implications — particularly when those implications require technical domain expertise to understand. A solid-state battery breakthrough published in Nature Energy is public information within seconds of publication. The connection to a specific mid-cap materials company whose patent portfolio directly enables that chemistry — and which is trading at a valuation that doesn't reflect it — takes days or weeks for the market to discover. This system makes that connection in minutes.
How we differ from comparable strategies
| Strategy | Approach | Our Differentiation |
|---|---|---|
| ARK-style thematic | Human analyst conviction, long-term holds | Systematic signal discovery, faster reaction, smaller cap focus |
| News sentiment quant | NLP on financial news | Primary sources (academic/patent) before news coverage |
| Factor investing | Value, momentum, quality screens | Innovation catalyst as entry trigger, not valuation ratios |
| Options flow trading | Smart money signal following | Options flow used as confirmation, not primary signal |
Expected return profile
Target return
30–80% annual return on deployed capital. The satellite sleeve is the return engine — a single 5x position in a correctly identified breakthrough company (WATT: +150%, RGTI: +74% in current paper portfolio) can dominate the annual P&L. The core sleeve targets 15–25% annually with lower volatility.
Volatility & drawdown
Higher than index funds — this is an active, concentrated, small-cap strategy. Maximum drawdown is managed by stop losses (5% core, 10% satellite), a daily circuit breaker (5% portfolio loss halts trading), and the cash-only enforcement that prevents leverage amplifying downside.
Human Oversight — Confirmation Mode
The system is designed to operate in confirmation mode by default during the live capital phase. This means:
Every trade proposed, not executed
When the system identifies a swap or new position, it generates a full investment case — sell rationale, buy thesis, historical win rates for the domain, current macro context — and sends it to the portfolio manager via Telegram. No trade executes without human approval.
AI-assisted decision making
The portfolio manager can interrogate the system in plain English: "Why are you selling this?", "What does the geopolitical context say about energy right now?", "Show me everything the database has on solid state batteries this week." Claude responds with full data access and can revise proposals on instruction.
Weekly investor reporting
Every Sunday: portfolio P&L, trades executed, signals found, exit decision quality review (were our sells correct?), strategy memo from the AI on what's working and what to adjust. Transparent, data-driven, auditable.
Override capability
High-conviction override signals — verified acquisition announcements, FDA approvals, government contracts above 85% confidence — can bypass the normal swap threshold and rotate immediately. The manager retains veto on all trades.
Beyond Listed Equities — Deal Flow for Family Offices
The signal pipeline does not stop at the public market boundary. FSI monitors non-public companies across the same 33 technology domains — surfacing private fundraises, patent filers, and crowdfunding raises before they become widely known. This layer is designed for family offices and sophisticated investors who have the capital and accreditation to act on early-stage deal flow.
The information asymmetry is larger in private markets
In listed markets, institutional coverage compresses the window between signal and price. In private markets, that window can be 18–36 months. A company filing a patent cluster in quantum error correction today may not attract serious institutional attention until a Series B two years from now. FSI identifies these companies at the patent and academic citation stage — corroborated by our 55,000-article intelligence corpus — and scores them through a private-market adapted viability framework before that window closes.
Companies House Monitor
Weekly scan of new share allotments filed at Companies House in SIC codes mapping to FSI's technology domains. A share allotment filing indicates a fundraise has occurred — often weeks before any press announcement.
UK Private CompaniesSEC Form D Monitor
US private placement notices are required within 15 days of a round closing. FSI's full-text EDGAR search surfaces filings containing domain keywords — identifying US deep tech raises as they happen, not months later via press.
US Private PlacementsEquity Crowdfunding
Active raises on Crowdcube and Seedrs are scanned weekly and matched against the domain taxonomy. Companies raising in regulated crowdfunding are accessible to non-accredited investors — relevant for standard subscribers as well as family office clients.
Crowdcube · SeedrsPatent Filer Identification
EPO patent filings are scanned for applicants in FSI's domains who do not yet appear in the public equity pipeline. A company filing multiple patents in a domain before listing is often the earliest identifiable signal of commercial intent.
EPO · Patent IntelligenceCrunchbase Funding Rounds
Recent funding rounds are pulled via Crunchbase and cross-referenced with FSI's domain keyword taxonomy. Stage, round size, and investor profile are included to provide deal context beyond the headline figure.
Seed · Series A · Series BArticle Corpus Corroboration
Every private company identified is cross-referenced against FSI's 55,000+ article database. Mentions in academic papers, patent citations, and specialist press provide independent corroboration of technology maturity before the company is formally assessed.
55k Article Intelligence DB| Factor | Weight | Rationale |
|---|---|---|
| Domain momentum | 20 pts | Is the company's domain generating accelerating signals in FSI's intelligence pipeline? |
| Stage appropriateness | 15 pts | Seed and Series A carry the highest asymmetry. Pre-IPO stages score lower — less upside remaining. |
| Source quality | 15 pts | SEC Form D (highest — legally mandated, accurate) through to crowdfunding scrape (lowest). |
| Patent activity | 15 pts | Patent filings in the domain are the strongest early signal of commercial intent and defensible IP. |
| Article corroboration | 15 pts | Mentions in FSI's 55k article corpus provide independent validation of technology credibility. |
| Funding amount | 10 pts | £500k–£10M is the target range. Too small may indicate pre-commercial. Too large means less asymmetry. |
| Keyword depth | 10 pts | Number of distinct domain keywords matched in company description and filing text. |
B2B Intelligence Product — Family Office & Fund Licence
The private market intelligence layer is available as a standalone add-on to the FSI B2B licence. Family offices and boutique funds receive a weekly deal flow briefing — domain-filtered, viability-scored, corroborated by the technology intelligence corpus, and accompanied by a one-paragraph Claude-generated analyst note on each identified company. This briefing is produced automatically every Sunday and is available immediately on licence activation. Unlike curated deal flow services, FSI's pipeline runs continuously — not just when a founder chooses to submit.
Pricing: Available as an add-on to the standard B2B licence (£4,000/month) or as a standalone private market intelligence subscription (pricing on application). Contact: admin@frontiersignalintelligence.com
Key Risks — Stated Plainly
A strategy that only presents upside is a strategy you shouldn't trust. These are the genuine risks.
| Risk | Nature | Mitigation |
|---|---|---|
| Model risk | The LLM may extract incorrect company associations or assign high confidence to weak signals. The system's filters are designed to catch this, but not perfectly. | 9-gate validation pipeline. Human confirmation on all trades. 30-day exit tracking to identify systematic errors. |
| Liquidity risk | Small-cap stocks can have thin order books. Stop losses may execute at significantly worse prices than set, particularly on gap-down opens following adverse news. | 100k daily volume minimum. Position sizing limited to avoid moving the market. Core/satellite split ensures most capital is in liquid names. |
| Concentration risk | 12 positions is a concentrated portfolio. A single position going to zero would represent an 8–12% portfolio loss. | Max 4 positions per sector. Stop losses on all positions. Circuit breaker halts trading at 5% daily loss. No leverage. |
| Early-stage strategy | The reinforcement learning system needs 50+ closed trades to activate. Self-calibration is in early stages. The system is improving but not yet fully calibrated. | Human oversight in confirmation mode during calibration period. Parameters start conservatively and adjust gradually with bounded step sizes. |
| Macro regime change | A risk-off macro environment (recession, rate hikes, credit crisis) reduces appetite for early-stage tech. The satellite sleeve would underperform significantly. | Geopolitical intelligence layer automatically reduces domain multipliers in risk-off signals. Core sleeve provides defensive anchor. Circuit breaker limits portfolio drawdown. |
| API / execution risk | Dependency on third-party APIs (broker, market data, LLMs). Outage of any component could delay trade execution or signal discovery. | Multiple fallbacks at every layer (GPT-4o → Groq, Alpaca → Trading212). Cloud server with auto-restart. Telegram alerting on failures. |
30-Day Go-Live Plan
| Week | Action | Status |
|---|---|---|
| Week 1 | Legal investment agreement drafted and signed. Switch to Trading212 with small test capital. | In progress |
| Week 2 | Run paper and live Trading212 in parallel. Verify all order types execute correctly. Compare signals. | Planned |
| Week 3 | Deploy 20% of capital (£20k). Monitor every trade manually. Establish weekly investor reporting cadence. | Planned |
| Week 4 | Deploy remaining capital. Bot operates autonomously in confirmation mode — all trades proposed to manager before execution. | Planned |
| Month 3+ | Migrate to Hetzner (4GB RAM, €4/mo) for production. L4 reinforcement learning activates at 50 closed trades. | Planned |